The hidden costs of AI projects

The cost of an AI project is not the one written on the quote. Not because suppliers hide rumors, but because a significant part of the work happens inside the company and is not billed by anyone: someone has to explain the process, fix the data, try the system, train colleagues.

This article lists items that are almost always missing, so you can budget them before you discover them in the works.

The six voices that are almost always missing

1. The internal time. It is the largest voice and the least considered. A project requires someone to explain how the process really works, not how it is written in procedures; provide real examples; test the system on real cases; give timely feedback. These are days of work by key people, who in the meantime don't do their jobs.

2. Data preparation. Records with duplicates, documents in multiple versions, free fields compiled in ten different ways. No AI system solves confusing data: it amplifies them.

3. Training. A few hours, low cost, huge impact. A tool that people don't understand is misused or circumvented.

4. Variants in progress. They arise from the exceptions not emerged in analysis. They are the most angry voice, and it is almost always avoidable with a serious mapping at the beginning.

5. Maintenance. Corrections, adjustments when changing a connected system, updates. Treated in how much it costs to maintain an AI system.

6. Evolution. A system that works is used more, and use generates demands. It's the cost of success, but it's still a cost.

Why Data Costs

In almost all companies, data exists but is not ready. Typical situations:

• The same customer presents three times with slightly different social reasons
• A note field where everything ended, because there wasn't the right field
• Documents in different folders, with multiple versions and no indication of what is worth
• Information that only exists in the head of those who have worked there for years
• Different nomenclatures between departments for the same thing

This work must be done and has a value that goes beyond the project: it also serves those who enter the company, reporting, continuity. But it needs to be budgeted, because it doesn't do itself and requires someone who knows the domain.

How to Avoid Variants

Variants in the works have an almost unique cause: exceptions discovered late. Prevention is trivial and hardly anyone applies it.

1. Before asking for a quote, write the process in ten lines.
2. Then ask the person who performs it, "What happens when it doesn't go as described?"
3. Write down each answer. It's the exceptions.
4. Deliver them to the supplier along with the request.
5. Ask that the quote explicitly state what exceptions are included.

Step 2 must be done with the people who work, not with those who coordinate them. The difference between the process described and the real process is exactly the space in which the variants are born.

Practical examples

Passive billing automation. Estimate of €3,000. Unscheduled costs: four days of administration to provide examples and try, two days to fix the supplier registry, half a day of training. The real cost is considerably higher, but the project works. If those days had not been given, the system would have come out calibrated on non-representative cases.

Customer Care Assistant. Estimate of €8,000. Unexpected cost: the sorting of the standard answers, which were scattered between emails, an old document and the memory of two people. Three weeks of part-time internal work. It was the most tiring and also the most useful piece: that base now also serves to form the new ones.

Project abandoned after six months. Well-built system, never adopted. No one had been trained, no one had a mandate to follow him, and at the first difficulty people returned to the method before. Cost: the entire investment, no return. It is the most common failure and has nothing to do with technology.

The Cost That No One Calls Cost: Adoption

There is an entry that does not appear in any quote and that determines whether everything else was spent well or thrown away: getting people to really use the system.

It's not technical training. It is the work of convincing those who have done that thing by hand for ten years that it is worth changing the method, of collecting the first objections without liquidating them, of correcting the things that do not actually work before they become the excuse to go back.

This work takes time of someone with internal authority, and needs to be expected. The companies that skip it find themselves in a precise and frequent situation: the system works, the supplier has delivered, and people continue to work as before using the system only when someone is watching.

1. Involve who will do the job already being analyzed, not on delivery.
2. Start with those who are favorable: a convinced colleague is worth more than a directive.
3. Gather objections in the first few weeks and quickly correct the real things.
4. Make the result visible: shows the numbers of before and after.

Step 1 is the most effective and is free. People who have helped define how a system works tend to use it; those to whom it is delivered tend to suffer it.

Voices at a glance

The first three lines are not billed by anyone and for this reason they disappear from evaluations. They are, however, the only ones that, if they are missing, cause a technically successful project to fail.

The price framework by project type is in how much AI costs in the company, the recurring part in how much it costs to maintain an AI system.

Mistakes to Avoid

• Not quantifying internal time. It is the largest and least visible voice.
• Assuming that the data is ready. They almost never are.
• Cutting training. It is the cut with the worst ratio of savings to damage.
• Not mapping exceptions. They become full-cost variants.
• Not planning for maintenance and evolution. They exist even when all is well.
• Choosing the lowest quote. Often it is the one that omitted the most voices.

How to apply it in the company

Take the quote you received and add three lines of your own:

• Internal days estimated for analysis, evidence and feedback, valued at company cost
• Time for data preparation, asking who handles them
• Annual maintenance budget, as a percentage of the project

The total you get is the real cost. It's higher than the one on the quote, but it's the number you can really decide on — and most importantly it's the number you can compare two different offers to in an honest way. The overall picture is in how much AI costs in the company and planning in AI budget for SMEs.

A final consideration concerning the relationship with the supplier should be added. A quote that explicitly includes analysis, testing, and assistance seems more expensive than one that omits them, but it is the one that is least likely to generate surprises. The difference between the two offers is not the price: it is where the unscheduled work ends, whether within the contract or within the days of your team.

Conclusion

The hidden costs of AI projects are not commercial tricks: they are real work that happens inside the company and that no one bills. People's time, data preparation, training, maintenance.

Budgeting them first changes two things: it makes comparisons between suppliers honest, and above all avoids the worst situation, the one where the project technically works but nobody uses it because there was no time to adopt it.

If you want to understand what processes in your company can be automated with AI, Giallo Studio can help you turn the problem into a real workflow.